Abstract:Fast and reliable solvers for optimal power flow (OPF) problems are attracting surging research interest. As surrogates of physical-model-based OPF solvers, neural network (NN) solvers can accelerate the solving process. However, they may be unreliable for ``unseen" inputs when the training dataset is unrepresentative. Enhancing the representativeness of the training dataset for NN solvers is indispensable but is not well studied in the literature. To tackle this challenge, we propose an OPF solver based on a physical-model-integrated NN with worth-learning data generation. The designed NN is a combination of a conventional multi-layer perceptron (MLP) and an OPF-model module, which outputs not only the optimal decision variables of the OPF problem but also the constraints violation degree. Based on this NN, the worth-learning data generation method can identify feasible samples that are not well generalized by the NN. By iteratively applying this method and including the newly identified worth-learning samples in the training set, the representativeness of the training set can be significantly enhanced. Therefore, the solution reliability of the NN solver can be remarkably improved. Experimental results show that the proposed method leads to an over 50% reduction of constraint violations and optimality loss compared to conventional NN solvers.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve the reliability and generalization ability of the neural network (NN) - based solver in the optimal power flow (OPF) problem. Specifically, although the traditional OPF solver based on the physical model is accurate, its computational efficiency is low. While the NN - based solver can accelerate the solution process, it may be unreliable when dealing with "unseen" inputs, especially when the training data set is not representative enough. Therefore, this paper proposes a neural network method combined with the physical model and introduces a data - generation technique worthy of study to enhance the representativeness of the training data set, thereby significantly improving the reliability and generalization ability of the NN solver. ### Main Contributions 1. **Designed a new neural network integrated with the physical model**: - This network consists of a traditional multi - layer perceptron (MLP) and an OPF model module. - The output includes not only the optimal decision variables of the OPF problem but also the degree of constraint violation. - By penalizing the degree of constraint violation in the loss function, the network can generate more reliable decision variables. 2. **Proposed a new data - generation method worthy of study**: - Based on the designed neural network, this method can identify samples that are not well - generalized within the input feasible region. - By iteratively applying this method and adding newly - identified worthy - of - study samples to the training set, the representativeness of the training set is gradually improved. - The results show that this method can significantly reduce constraint violations and optimality losses. Compared with the traditional NN solver, the constraint violations are reduced by more than 50%. ### Method Overview 1. **Framework**: - Training set initialization: Initialize the training set by random sampling. - Training the neural network: Train the physical - model - integrated NN on the current training set. - Identifying data worthy of study: Identify new worthy - of - study samples according to the output of the current NN. - Updating the training set: If data worthy of study is identified, add it to the training set and return to the second step; otherwise, output the current NN. 2. **Neural network integrated with the physical model**: - **Traditional NN module**: Use the traditional MLP to fit the mapping from load to optimal decision variables and adjust the output through the activation function to meet the constraint conditions. - **Physical model module**: Receive the output of the previous module, calculate the optimal power generation and the corresponding degree of constraint violation. 3. **Loss function**: - The loss function consists of three parts: voltage prediction error, power prediction error, and the degree of constraint violation. - By combining these three parts, the fitting accuracy is improved and the network is guided to generate more reliable solutions. 4. **Data - generation worthy of study**: - Identify samples worthy of study by maximizing the degree of constraint violation. - Ensure that the identified samples are feasible in the original OPF problem. - Use gradient back - propagation and automatic differentiation techniques to efficiently find a large number of samples worthy of study. ### Experimental Results The experimental results show that the proposed method performs excellently in terms of reliability, optimality, and computational efficiency. Especially when dealing with the OPF problem in high - dimensional input spaces and under complex constraint conditions, it is significantly superior to the traditional NN solver. In conclusion, this paper effectively solves the reliability problem of the NN - based OPF solver when dealing with "unseen" inputs by combining the physical model and the data - generation technique worthy of study, providing strong support for the efficient operation of modern power systems.

Optimal Power Flow Based on Physical-Model-Integrated Neural Network with Worth-Learning Data Generation

OptNet-Embedded Data-Driven Approach for Optimal Power Flow Proxy

Optimal Power Flow Using Graph Neural Networks

Unsupervised Optimal Power Flow Using Graph Neural Networks

DeepOPF+: A Deep Neural Network Approach for DC Optimal Power Flow for Ensuring Feasibility

Learning to Solve the AC Optimal Power Flow via a Lagrangian Approach

DeepOPF: Deep Neural Network for DC Optimal Power Flow

Graph Neural Network-Accelerated Network-Reconfigured Optimal Power Flow

Advancing Hybrid Quantum Neural Network for Alternative Current Optimal Power Flow

AC Optimal Power Flows: Combining Lagrangian Dual and physics-Guided Graph Neutral Networks

Topology-aware Graph Neural Networks for Learning Feasible and Adaptive ac-OPF Solutions

Input Convex Neural Network-Assisted Optimal Power Flow in Distribution Networks: Modeling, Algorithm Design, and Applications

DeepOPF-U: A Unified Deep Neural Network to Solve AC Optimal Power Flow in Multiple Networks

Model-Informed Generative Adversarial Network (MI-GAN) for Learning Optimal Power Flow

Constraints and Variables Reduction for Optimal Power Flow Using Hierarchical Graph Neural Networks with Virtual Node-Splitting

Optimal Power Flow in Highly Renewable Power System Based on Attention Neural Networks

Optimal Power Flow Estimation Using One-Dimensional Convolutional Neural Network

A Neural Network Approach to Physical Information Embedding for Optimal Power Flow

Unsupervised Learning for Solving AC Optimal Power Flows: Design, Analysis, and Experiment

A Trustable Data-Driven Optimal Power Flow Computational Method With Robust Generalization Ability

Robust Training for AC-OPF (Student Abstract)